Improving field boundary delineation in ResUNets via adversarial deep learning

Field boundary data is often required to access digital agricultural services and tools that assist with field-level assessment and monitoring. In addition, policy-makers and researchers need field boundaries to accurately assess food security and impacts on climate change. Thus, scalable and efficient automatic field boundary detection algorithms on satellite images have direct, important implications for many stakeholders. Deep learning is one approach that has been successfully applied in recent years to field boundary detection. Qualitatively however, these boundaries are often broken or malformed, necessitating a dependence on fine-tuned post-processing methods with arbitrary thresholds obtained through trial-and-error. Prior work has explored various architectures for predicting field boundaries, but little has been done beyond traditional supervised learning regimes. Thus, in this work, we propose a new approach to improving field boundary prediction by using an adversarial training framework. In particular, we investigated the effects of training a ResUNet model (a standard fully convolutional network architecture) as a generator in a traditional generative adversarial network (GAN) setup, on 30 meter resolution satellite imagery from 2017 over the state of Illinois. We then explored whether or not our methods can be transferred to label-scarce regions in Brazil. Overall, our results showed that adversarial training substantially improved boundary quality and performance, but had a lesser effect when transferred to unseen, low-data agricultural landscapes. Based on these findings, we conclude that adversarial training is a promising way to improve boundary quality during prediction time, and we suggest several ideas for future improvements that may make adversarial training more viable in transfer learning.